A New Baseline Estimation Method Applied to Arabic Word Recognition

We analyse in this paper the impact of different baseline identification approaches in the case of single word recognition. We show that classical baseline identification approaches using horizontal projection histograms may fail in detecting accurately the baseline of short words, impacting the overall processing chain and inducing errors. From this observation, we propose a novel approach based on stochastic models able to propose probable baseline regions from characters features. Once the most probable baseline region is detected, we fine tune the position of the baseline with an horizontal projection histogram. We ran our experiments in the case of a printed word recognition task using the APTI database and observed a significant increase of performance. Keywords-HMM; GMM; arabic recognition; baseline;